Learning mobility pro les of citizens can play a crucial role in many applications, including tra c demand estimation, urban planning or personalized advertising. In this paper we demonstrate a framework for building and constantly readjusting mobility pro les using smart phone data coupled with manual user input and personalised discrete choice models. The methods are applied as weather warning service supporting the daily mode choice decisions of users of the system by supplying personalised information based on their mobility pro le and current weather conditions. Since it is well known that weather conditions in uence the tra c demand and the modal split of transport modes, the framework can also further the understanding of mobility patterns and their variability due to weather or tra c events.
Learning mobility pro les of citizens can play a crucial role
in many applications, including tra c demand estimation,
urban planning or personalised advertising. Progress has
been made both in the area of mobility pro les as well as
the estimation of travel demand. For Mobility pro les this
is done by applying data analysis techniques to data sources
like mobile phone data or data collected using smart phones
(GPS, accelerometer data). In [
Another approach to reach mobility pro les is that of
applying mobility diaries. While the sample sizes are
considerably smaller than in the case of mobile phone data the rise
of smart phone availability makes the collection and
enrichment of the data through mode detection easier and enables
the collection of larger samples. In [
To learn more about the long term behaviour of people data needs to be collected more long term and new methodologies for the analysis of the data need to be developed. In this paper we demonstrate some missing steps to get from collected long term data to mobility pro les and personalised mode choice models. This framework can be applied to build and constantly readjust mobility pro les of users with data collected with smart phones (GPS-data, accelerometer data) coupled with some manual user input. Finally, this framework is applied as personalised weather warning service. The system supplies personalised information based on the mobility pro le enhanced by individualized discrete choice models and current weather conditions to the user to support their daily mode choice decisions. Since it is well known that weather conditions in uence the tra c demand and the modal split of transport modes, the framework can also further the understanding of mobility patterns and their variability due to weather or tra c events.
The paper will rst present the data collection and the learning algorithms for the mobility pro les in section 2 before describing the personalisation of the mode choice models. In section 3 the nal weather warning system will be presented. The nal section will give some conclusions and an outlook to future work.
As a basis for the learning process of the mobility pro le,
GPS and other cell phone data (e.g. accelerometer data, cell
position) were collected on the users daily trips. For each
recorded trip there is a data preprocessing step before the
main learning process. This cleans the data and detects the
transport mode by transforming the collected GPS,
acceleration and cell phone signal data, following the methods of
[
Within a three week long experiment users were asked to record their daily trips with a smart phone application, which was developed especially for this very experiment. The application was easy to handle and designed to require minimum user interaction. Four steps had to be done per trip: (1) start the trip by opening the application and starting the recording, (2) state the travel purposes, (3) end the trip by opening the application again and nishing the recording, (4) adjust the detected modes instantaneously on the device. In between steps three and four trips were automatically cut into uni modal stages and the mode was detected as described above.
The pro le learning algorithm consists of three steps; (1) identify points of routine (POR), (2) identify routine trips (RT) connecting the PORs, and (3) learning about the user's travel behaviour by modelling their choice situations (see section 2.3). On a regular basis the user pro les are re ned by applying the learning algorithm including newly collected records.
For every trip the GPS positions are recorded, so the origin and destination can be de ned as the rst and last GPS points of a trip. Consolidating these by an agglomerative hierarchical cluster method the PORs are de ned as the clusters' centers. Since the main emphasis is detecting travel patterns and routine trips, only clusters containing a minimum number of four points are considered as POR, whereas smaller ones may build another one as soon as more trips start or end in the same region. The size of a cluster was de ned by the clustering height, that is the agglomeration was stopped at a height of 500 meters.
The cluster analysis was performed in R, applying agnes
[
Once a pro le includes at least two PORs, the connecting trips are aggregated to formulate routine trips. Each RT contains information about the usual travel time, main transport means and earliest and latest departure time for two PORs. A trip's main transport mean is de ned as the one chosen for the biggest distance. In this work the following main transport means are considered: walk, bicycle, public transport and car. Consequently for one RT up to four alternatives can be learned. Knowing the available alternatives for a RT is crucial for formulating: the choice set in choice model estimation (see section 2.3), and the weather warning service (see section 3).
For the algorithm of personalising mode choice models we
follow the methodology of [
Uni(m) = Xnim + nim where Xnim is a vector of decision variables for using mode m for that trip, N ( ; ) are the normally distributed parameters from the mixed logit model with mean and covariance Matrix and is a extreme value distributed random error.
To get to the personalised models the choice situations of the user are used by 1. draw a sample R of size R from N ( ; ), where R is a large number (R = 25000 in our implementation) 2. calculate the personalised parameters as
R = X (r) r=1
P (ymjUni(m); (r)) PR
s=1 P (ymjUni(m); (s)) where ym is one if m is the chosen mode and zero otherwise and P (ymjUni(m); (b)) = Pekxepx(Up(nUi(nmi()k))) , where the sum in the denominator is over all routes observed up to this point for the user.
An example of the changes in parameter values for one user can be seen in Figure 2. The user is an avid bike rider. As a result the alternative speci c constant for bike rises quickly to a value between 2:5 and 3. This raises the likelihood that the model predicts the mode bike for that user. One can also see that at low temperatures without rain the person is more likely to walk than the average user and that biking becomes more likely for higher temperatures without rain. For some parameter values it can be seen that they stay at constant values until a trip of that category is observed before there are changes in that parameter value.
For applying the learning algorithm in a real world experiment users were provided with a smart phone application that collects and transmits trip data and receives personalised weather messages. The main purpose was to provide the users with information whenever it would be helpful for their decision process, rather than sending redundant messages, so that the user stayed motivated to continue using the app and recording trips. Finally, this would sharpen the mobility pro le and result in more relevant messages to prevent users from experiencing unpleasant weather situations for their chosen modes, i.e. prevent them from choosing a mode that they would not have chosen with perfect information. Based on the personalised choice model the user's utilities for di erent transport modes were calculated, both for the current weather forecast and optimal weather conditions (no rain and 15-25 degrees C). For each RT only utilities of modes are compared, that the user has already recorded for this RT. In case of expected behaviour change the user was informed about the weather forecast and a favoured transport mode choice, otherwise a generic weather message was sent. Figure 3 depicts a personalised message, both on the lock screen of the smart phone and within the application. The experiment showed that the learning algorithm performed as expected. It learned the users' PORs and RTs as demanded and adjusted the personalised mode choice parameters continuously. Ten users took part and recorded on average 40 trips each. Two users left the experiment in an early stage, so that no PORs could be recognised. For each of the remaining eight users the home location could be identi ed. Further the algorithm found 6 education related PORs, 3 work PORs and one for leisure purpose. For 6 users the algorithm could learn on average 3.2 RTs, whereas for each of these one to four alternatives (walk, bicycle, car, public transport) with di erent main transport means were identi ed.
In this project algorithms were developed that create and update a user mobility pro le, estimate personalised modechoice models and nally provide users with weather and mode information supporting their daily mobility decisions. The system was tested in a eld test showing promising results. Due to the stable nice weather during the experiment time span of March 2015 as well as the limitations of the underlying mobility survey data-set that was collected in late spring, learning the weather sensitivity of the users' mobility behaviour was challenging. However the individual level parameters approach learning about sensitivities to weather events can be incorporated into the models quickly after observing such an event. Furthermore due to the short test period, the mobility pro les lack some grade of detail. Further steps will deal with developing the algorithm to send weather messages dependent on the weekday. Therefor a longer data collection and learning period is necessary. Consequently for sharpening the pro le also older trips could be skipped for keeping the input information for pro ling up to date. For a more precise evaluation method user interaction would be required. This could either be achieved by a special survey or directly integrated in the smart phone application.
Furthermore, to improve the system, automated start and stopping of the data collection would limit the reliance on the user's willingness to record trips and would improve proling and would better enable the next step of suggesting alternative routes rather than just alternative modes. Lastly, the integration of other data like tra c events would improve the system further.
This research was partially supported by the research project Wetter-PROVET. The project was funded through the FFG - Austrian Research Promotion Agency program Ways2Go by the Austrian Ministry for Transport, Innovation and Technology (BMVIT). We also thank the project partners Fluidtime Data Services GmbH and UBIMET GmbH.